Predictive modeling of motor symptom severity and stage classification in Parkinson's disease using machine learning methods with selected multiple serological biomarkers.
Predictive modeling of motor symptom severity and stage classification in Parkinson's disease using machine learning methods with selected multiple serological biomarkers.
Where did the research take place?
The study site has not been established. Author addresses may differ from where the research occurred.
Guangzhou, CN · Author affiliation
The Department of Neurology; Guangdong Provincial Key Laboratory of Major Obstetric Diseases; Guangdong Provincial Clinical Research Center for Obstetrics and Gynecology, The Third Affiliated Hospital, Guangzhou Medical University, Guangzhou, 510150, PR China.Location evidence
Zhengzhou, CN · Author affiliation
Henan Key Laboratory of Chronic Disease Prevention and Therapy & Intelligent Health Management, Henan Institute of Interconnected Intelligent Health Management, Zhengzhou, 450045, PR China. lianglysky@gmail.com.Location evidence
A plain-language reading has not been prepared for this paper yet.
Original abstract
BACKGROUND: Parkinson's disease (PD) affects approximately 1% of individuals aged 60 and older, with prevalence projected to double by 2030 amid global aging. Traditional assessments using the Unified Parkinson's Disease Rating Scale (UPDRS) rely heavily on subjective clinical evaluation, producing inconsistent and non-reproducible outcomes. Accumulating evidence demonstrates that PD involves systemic inflammatory responses, making serological biomarkers promising candidates for objective severity assessment in routine clinical settings METHODS: We analysed serological data from 2,614 individuals to investigate correlations with motor symptom severity and Hoehn and Yahr (H&Y) staging. Thirty clinically relevant biomarkers were selected, including inflammatory markers (NLR, LMR, SII) and biochemical indicators (albumin, ALT, uric acid). We developed multiple machine learning models—XGBoost, CatBoost, Random Forest, and Decision Tree—for H&Y classification and MDS-UPDRS motor score prediction (total and itemized). Feature selection was conducted using XGBoost importance rankings. Additionally, we constructed a deep neural network incorporating adaptive feature selection and learning rate adjustment strategies to evaluate deep learning performance with the same feature set. RESULTS: Among traditional machine learning models, XGBoost achieved the highest performance for H&Y classification (accuracy: 0.816, weighted F1-score: 0.813), followed by CatBoost (accuracy: 0.812, F1-score: 0.811) and Random Forest (accuracy: 0.803, F1-score: 0.801). XGBoost also demonstrated superior performance in regression tasks for predicting motor scores. The DNN model significantly outperformed all classifiers, attaining accuracy of 0.885 and weighted F1-score of 0.881 for H&Y staging. Feature importance analysis revealed that inflammatory markers—particularly neutrophil-to-lymphocyte ratio (NLR) and lymphocyte-to-monocyte ratio (LMR)—emerged as critical predictors across all tasks. Removal of these markers resulted in notable performance decline, confirming their essential role in disease severity assessment and highlighting the relevance of systemic inflammation in PD progression. CONCLUSION: Serological biomarkers, particularly inflammatory markers, can serve as effective non-invasive indicators for assessing PD severity and staging. Machine learning and deep learning algorithms, enhanced by adaptive optimization strategies, significantly improve the accuracy and objectivity of clinical evaluations. These findings establish a foundation for developing scalable, accessible diagnostic tools that support early PD detection, progression monitoring, and treatment planning in routine clinical practice.